Device variability of Josephson junctions induced by interface roughness Preprint: AIP/123-QED Yu Zhu, Félix Beaudoin, Hong Guo Abstract As quantum processors scale to large qubit numbers, device-to-device variability emerges as a critical challenge. Superconducting qubits are commonly realized using Al/AlO x {}_{\text{x}} /Al Josephson junctions operating in the tunneling regime, where even minor variations in device geometry can lead to substantial performance fluctuations. In this work, we develop a quantitative model for the variability of the Josephson energy E J E_{J} induced by interface roughness at the Al/AlO x {}_{\text{x}} interfaces. The roughness is modeled as a Gaussian random field characterized by two parameters: the root-mean-square roughness amplitude σ \sigma and the transverse correlation length ξ \xi . These parameters are extracted from the literature and molecular dynamics simulations. Quantum transport is treated using the Ambegaokar–Baratoff relation combined with a local thickness approximation. Numerical simulations over 5,000 5,000 Josephson junctions show that E J E_{J} follows a log-normal distribution. The mean value of E J E_{J} increases with σ \sigma and decreases slightly with ξ \xi , while the variance of E J E_{J} increases with both σ \sigma and ξ \xi . These results paint a quantitative and intuitive picture of Josephson energy variability induced by surface roughness, with direct relevance for junction design. I Introduction Superconducting circuits are among the most promising platforms for scalable quantum computing, combining lithographic manufacturability with strong, low-loss nonlinearity from the Josephson effect 3 . At the heart of widely used transmon-style devices is an Al/AlO {}_{\text{x}} /Al Josephson junction shunted by a capacitance, forming a weakly anharmonic quantum oscillator whose lowest two eigenstates \left|0\right\rangle and \left|1\right\rangle serve as the computational basis 7 . In this picture, the qubit transition frequency is set primarily by the Josephson energy E_{J} and charging energy E_{C} , so any process-dependent change in the junction’s critical current (and hence E_{J} ) directly shifts the qubit spectrum and operating point. State-of-the-art superconducting quantum processors now contain on the order of 10^{2} to 10^{3} qubits, exemplified by IBM’s 1,121 -qubit “Condor” superconducting processor 5 . However, fault-tolerant quantum computation for societally relevant tasks remains far beyond this scale. For example, detailed resource estimates for Shor’s algorithm indicate that factoring a 2048-bit RSA integer would require on the order of tens of millions of noisy physical qubits in a fast-runtime scenario (e.g., \sim 20 million noisy qubits for an 8-hour attack under surface-code assumptions) 6 . This large gap motivates careful attention to device yield and uniformity as superconducting platforms scale toward much larger systems. Josephson junction variability becomes a central challenge at scale because junctions operate in the tunneling regime, where the critical current I_{c} depends exponentially on microscopic barrier details; small perturbations—such as local Al/AlO {}_{\text{x}} thickness fluctuations, effective barrier-height variations, or contamination atoms—can produce significant shifts in E_{J} , qubit frequency, and gate fidelity. Experimentally, substantial efforts over the past two decades have focused on improving junction reproducibility and understanding microscopic sources of variation, including wafer- and chip-scale resistance/critical-current uniformity studies 8 ; 14 , process-optimization of shadow evaporation/oxidation 14 ; 10 , mitigation strategies for critical-current fluctuations 15 , alternating-bias assisted annealing 13 , and direct microscopy of Al/AlO {}_{\text{x}} barrier microstructure and thickness distributions 17 ; 4 . In this work, we focus on variability in E_{J} induced by Al/AlO {}_{\text{x}} interface roughness, and present a quantum-mechanical modeling framework that combines quantum transport theory with Gaussian random field models and molecular dynamics simulations to connect microstructural disorder to junction-level Josephson energy statistics. II Interface roughness model The rough interface between the aluminum (Al) leads and the aluminum oxide (AlO {}_{\text{x}} ) layer is modeled as a Gaussian random field with a specialized power sepctral density. Denote the interface height by h\left(x,y\right) , where x and y are the tranverse coordinates. It is assumed that \left\langle h\left(x,y\right)\right\rangle=0, (1) and \left\langle h\left(x,y\right)h\left(0,0\right)\right\rangle=\sigma^{2}e^{-\frac{x^{2}+y^{2}}{\xi^{2}}}, (2) where \sigma is the root-mean-square (RMS) of the interface height along the transport direction, and \xi is the transverse correlation length. Given \sigma and \xi , a rough interface can be generated by filtering white noise in Fourier space using the corresponding power spectral density, followed by an inverse Fourier transform. Two generated rough interfaces are shown in Fig. 3 (a) and (b). The transverse correlation length \xi is set by the characteristic lateral length scale of microstructural features in the Al leads—most notably the Al grain size that imprint thickness variations onto the AlO {}_{\text{x}} barrier. In Al/AlO {}_{\text{x}} /Al tunnel junction stacks characterized by TEM, the Al electrodes are commonly observed to be polycrystalline with columnar grains on the order of a few tens of nanometers. For example, Liu et al . report crystalline columnar grains of roughly \sim 20 nm \times 30 nm in qubit Josephson junction electrodes 4 . A systematic TEM study by Nik et al. further shows that the average lateral grain size depends strongly on the Al film thickness: for Al films of 15 nm and 60 nm thickness on oxidized Si, the mean grain size is reported to be on the order of \sim 38 nm and \sim 92 nm, respectively, illustrating that thicker Al films tend to develop larger grains 12 . Under other deposition conditions, Fritz et al. report that the lower Al layer can exhibit a broad grain size distribution, including grains below \sim 50 nm, many grains in the \sim 50 nm-150 range, and grains extending up to hundreds of nanometers ( \sim 900 nm) and even \sim 1 \mu m in lateral size 4 . These experimental observations suggests choosing \xi to be comparable to the typical lateral grain size for the specific Al growth recipe used in the junction fabrication. The RMS height \sigma is set by the vertical amplitude of thickness/height fluctuations, primarily controlled by the Al surface roughness and oxidation/growth nonuniformity. Zeng et al. directly measured the local AlO {}_{\text{x}} thickness in Al/AlO {}_{\text{x}} /Al junctions using atomic-resolution STEM and fit the thickness histogram with a Gaussian. They report mean thicknesses 1.66–1.88 nm and standard deviations 0.326–0.372 nm in three samples 17 . Fritz et al. report AlO {}_{\text{x}} thickness variation values ranging from 0.29 nm down to 0.11 nm depending on growth conditions 4 . These experiments suggest a representative range of \sigma \sim 0.1–0.3 nm for high-quality AlO {}_{\text{x}} tunnel barriers. Since \sigma is much smaller than \xi and more difficult to characterize experimentally, we also perform molecular dynamics (MD) simulations to estimate \sigma . The initial simulation cell contains an Al surface, O {}_{\text{2}} molecules, and a vacuum region. In total, the system has 4,380 atoms, comprising 2,880 Al atoms and 750 O {}_{\text{2}} molecules, as illustrated in Fig. 1 (a). The Al surface is modeled as an fcc(100) slab consisting of 20 atomic layers, with 144 atoms per layer and a 40 Å vacuum region. The O {}_{\text{2}} molecules are placed at random positions with random orientations in a 10 Å thick region between the Al surface and the vacuum. MD simulations are carried out in the NVT ensemble at a fixed temperature 300 K and a fixed volume 34.17 Å \times 34.17 Å \times 78.26 Å, with periodic boundary conditions in the transverse directions. Interatomic forces and energies are computed using CHGNet, a charge-informed graph-neural-network machine learning interatomic potential pretrained on the Materials Project Trajectory dataset 2 . MD simulations are performed for 3 ps with a time step of 1 fs. During the first \sim 2 ps, O {}_{\text{2}} molecules react with the Al surface and form an amorphous AlO {}_{\text{x}} layer with a thickness about 1 nm, as shown in Fig. (b). Then the oxide growth saturates and the remaining O {}_{\text{2}} molecules diffuse into the vacuum region. After generating the Al/AlO {}_{\text{x}} interface, we remove Al lead atoms and O {}_{\text{2}} molecules from the system to expose the AlO {}_{\text{x}} surface, as shown in Fig. 1 (c). To extract a smooth height profile, we employ a rolling-ball procedure: a probe sphere of radius R=3 Å is rolled over the AlO {}_{\text{x}} surface on the Al side to define a continuous surface envelope. The RMS of the surface \sigma is obtained as 0.97 Å, 0.76 Å, 0.81 Å in three independent MD simulations. We note that \sigma values extracted from our MD simulations are smaller than the experimentally reported values. This discrepancy is likely due to the idealized geometry used in the simulations: the Al surfaces are assumed to be atomically flat, whereas real devices typically exhibit additional Al surface roughness that increases the overall height fluctuations. Figure 1: MD simulation of AlO {}_{\text{x}} growth. (a) At t=0 , fcc(100) Al surface and O {}_{\text{2}} molecules. (b) At t=3 ps, oxidized Al surface and O {}_{\text{2}} molecules. (c) AlO {}_{\text{x}} layer extracted from (b). III Quantum transport model In a Josephson junction, the Josephson energy E_{J} is related to the critical current I_{c} by E_{J}=\frac{\hbar}{2e}I_{c}, (3) where \hbar is is the reduced Planck constant and e is the elementary charge. The critical current can be calculated by solving the Bogoliubov–de Gennes (BdG) equation \left(\begin{array}[]{cc}H_{0}\left(\mathbf{r}\right)-E_{F}&\Delta\left(\mathbf{r}\right)\\ \Delta^{\ast}\left(\mathbf{r}\right)&-\left(H_{0}\left(\mathbf{r}\right)-E_{F}\right)\end{array}\right)\left(\begin{array}[]{c}u\left(\mathbf{r}\right)\\ v\left(\mathbf{r}\right)\end{array}\right)=E\left(\begin{array}[]{c}u\left(\mathbf{r}\right)\\ v\left(\mathbf{r}\right)\end{array}\right), (4) where H_{0}\left(\mathbf{r}\right)=-\frac{\hbar^{2}}{2m}\nabla^{2}+V\left(\mathbf{r}\right) is the single-particle Hamiltonian, E_{F} is the Fermi energy, \Delta\left(\mathbf{r}\right) is the superconducting pair potential, u\left(\mathbf{r}\right) and v\left(\mathbf{r}\right) are the electron- and hole-like components of the quasiparticle wavefunction. There are two equivalent theoretical approaches to calculate the supercurrent from the model in Eq. ( 4 ). The first approach is the scattering wave function method 1 , which provides a clear picture about physical processes such as Andreev reflections and Andreev bound states. The second approach is the Green’s function method 16 , which offers a unified formalism for general cases and allows effective computational optimization to deal with Andreev bound states 19 . However, applying these methods directly to a realistic Josephson junction is not computationally affordable since the size of transport region is as large as 100 nm \times 100 nm \times 1 nm. The Fermi wavelength of Al, on the other hand, is around 3.6 Å, estimated using the Fermi energy 11.6 eV. Consequently, the real space resolution should be around 1 Å, resulting in a 10^{3}\times 10^{3}\times 10 real space grid which is too large for numerical simulations. To tackle the computational challenge, we adopt two approximations to simplify the transport model. The first approximation is to use the Ambegaokar–Baratoff (AB) relation I_{c}=\frac{\pi\Delta}{2eR_{N}}, (5) which relates the critical current I_{c} to the normal resistance R_{N} . The calculation of R_{N} is much simpler than I_{c} because it only requires an evaluation of the transmission coefficient at the Fermi energy rather than an integration over the entire energy range. Furthermore, the system size is reduced by half since only the electron-like quasiparticles need to be considered. To assess the accuracy of the AB relation for Josephson junctions, we compare the critical current obtained by solving Eq. ( 4 ) with that obtained from the AB relation for a Josephson junction with a square tunnel barrier of height 1.1 eV. As shown in Fig. 2 , the results of the two methods agree in high accuracy over a wide range of junction thickness, justifying the use of the AB relation in this study. Moreover, the exponential dependence of E_{J} on the tunnel barrier thickness d confirms that transport occurs in the tunneling regime. Figure 2: Josephson energy (log scale) as a function tunnel barrier width, calculated by solving the BdG equation and using the AB relation. The second approximation is to separate the transport direction from the transverse directions. Because the junction operates in the tunneling regime, transport is predominantly governed by the local barrier profile along the transport direction, and transverse coupling is negligible. We therefore partition the junction into many independent subsystems in the transverse plane and approximate the total normal conductance as the sum of the conductances of these subsystems. In the continuum limit, this sum becomes an integral of the conductance density over the transverse coordinates. Here, the conductance density is defined as the conductance per unit area for a three-dimensional junction with a uniform cross section in the transverse directions. This local thickness approximation is justified by comparing the relevant length scales of the system. Four characteristic lengths are involved: (1) The Fermi wave length in the Al leads \lambda_{F}=\frac{2\pi}{k_{F}}=\frac{2\pi}{\sqrt{\frac{2mE_{F}}{\hbar^{2}}}}\approx 3.6\text{ \AA,} where E_{F}=11.6 eV the Fermi energy of Al bulk. (2) The wave function decaying length in the tunnel barrier \lambda_{D}=\frac{1}{\varkappa}=\frac{1}{\sqrt{\frac{2mU_{\text{barrier}}}{\hbar^{2}}}}=1.9\text{ \AA,} where U_{\text{barrier}}=1.1 eV is the effective tunnel barrier height of the Al/AlO {}_{\text{x}} interface. (3) The RMS of interface roughness along the transport direction: \sigma \sim 0.1 nm. (4) The correlation length of interface roughness in the transverse directions: \xi\sim 10 nm. The change of the tunnel barrier thickness over a lateral distance \lambda_{F} is estimated as \frac{\lambda_{F}}{\xi}\sigma . Since \frac{\lambda_{F}}{\xi}\sigma\ll\lambda_{D}. the tunnel barrier can be viewed as an ensemble of locally uniform tunnel barriers, which justifies the local thickness approximation. IV Results The interface roughness model and quantum transport framework are combined to study the variability of the Josephson energy E_{J} in Al/AlO {}_{\text{x}} /Al junctions. The junction is modeled as two free-electron metallic leads separated by a three-dimensional tunnel barrier. The tunnel barrier is characterized by a height U\sim 1 eV, a width d\sim 1 nm, and a cross section L_{x}\times L_{y}=200 nm \times 200 nm, consistent with Al/AlO {}_{\text{x}} interface. The leads are characterized by a Fermi energy E_{F}=11.7 eV and a superconducting gap \Delta=0.2 meV, consistent with bulk Al. The interface roughness between the two leads and the tunnel barrier is described by the Gaussian random field model in Eqs. ( 1 , 2 ) with parameters RMS amplitude \sigma and transverse correlation length \xi . For simplicity, it is assumed that \sigma and \xi are identical for the two Al/AlO {}_{\text{x}} interfaces in the Josephson junction. The resulting local barrier thickness is obtained by adding the two interfacial height profiles to the nominal thickness d , thus defining the spatially varying tunnel-barrier profile used in the transport calculation. To be more specific, let us choose the tunnel barrier parameters as d=1.0 nm and U=1.1 eV, and set the interface roughness parameters as \sigma=0.85 Å and \xi=10 nm. Fig. 3 (a) and (b) are the colormaps of the resulting interface roughness, showing fluctuations at the scale of transverse correlation length \xi . Fig. 3 (c) is the normalized histogram of the calculated E_{J} over 5,000 Josephson junctions with different tunnel barrier profiles. It is very interesting that the distribution of E_{J} is not a Gaussian but skewed toward larger values. This feature can be understood as a result of the tunneling physics. Suppose the effective tunnel barrier width is normally distributed and since E_{J} has an exponential dependence on the tunnel barrier width, the distribution of E_{J} must follow a log-normal distribution, and the corresponding probability density function is p\left(E_{J}>0\right)=\frac{1}{E_{J}\sigma_{J}\sqrt{2\pi}}\exp\left[-\frac{\left(\log E_{J}-\mu_{J}\right)^{2}}{2\sigma_{J}^{2}}\right], (6) where \mu_{J} and \sigma_{J} are the mean and standard deviation of \log E_{J} . Consequently, the mean and variance of E_{J} are \mathbb{E}\left[E_{J}\right]=e^{\mu_{J}+\sigma_{J}^{2}/2}, and \operatorname{Var}\left(E_{J}\right)=\left(e^{\sigma_{J}^{2}}-1\right)e^{2\mu_{J}+\sigma_{J}^{2}}, respectively. These expressions show why the distribution of E_{J} over the ensemble of the Josephson junctions becomes increasingly skewed and broadened as the underlying fluctuations grow. Fitting the histogram with Eq. ( 6 ), we obtain E_{J}/h=20.447\pm 2.326\text{ GHz.} (7) For transmons, the qubit transition frequency is determined by E_{C} and E_{J} as 7 \frac{\omega_{01}}{2\pi}=\frac{1}{h}(E_{1}-E_{0})\approx\frac{1}{h}(\sqrt{8E_{C}E_{J}}-E_{C}). Therefore, the standard deviation of \frac{\omega_{01}}{2\pi} is related to that of E_{J} by \Delta\frac{\omega_{01}}{2\pi}=\frac{1}{h}\sqrt{\frac{2E_{C}}{E_{J}}}\Delta E_{J}. Using the mean and standard deviation of E_{J} in Eq. ( 7 ) and assuming E_{C}/h=0.25 GHz, we obtain the mean and standard deviation of \omega_{01}/2\pi as \omega_{01}/2\pi=6.145\pm 0.364\text{ GHz.} (8) Figure 3: (a) and (b): Colormaps of Al/AlO {}_{\text{x}} interface roughness. (c): Histogram of E_{J} over 5,000 Josephson junctions. The orange line is a fit to Eq. 6 . Next, we investigate the dependence of the E_{J} variability on the roughness parameters \sigma and \xi . Table 1 lists the calculated mean and standard deviation of E_{J} as functions of \sigma and \xi , while fixing the other device parameters at U=1.1 eV, d=1.0 nm, \Delta=0.2 meV, E_{F}=11.7 eV. Two main trends are observed. (1) With increasing \sigma , both \mathbb{E}\left[E_{J}\right] and \operatorname{Var}\left(E_{J}\right) increase. \operatorname{Var}\left(E_{J}\right) increases because the amplitude of thickness fluctuations scales with \sigma , leading to stronger sample-to-sample variations in E_{J} . The rise in \mathbb{E}\left[E_{J}\right] reflects the exponential sensitivity of tunneling to the local barrier thickness: the total conductance is dominated by rare, anomalously thin regions, which become thinner and more influential as \sigma grows. (2) With increasing \xi , \mathbb{E}\left[E_{J}\right] decreases slightly, whereas \operatorname{Var}\left(E_{J}\right) increases markedly, and both trends saturate once \xi becomes comparable to the lateral size of the junction. The increase of \operatorname{Var}\left(E_{J}\right) can be understood as a consequence of reduced self-averaging. The junction may be viewed as comprising many parallel tunneling channels, each characterized by a correlation area of order \xi^{2} . A larger \xi implies fewer statistically independent channels, which in turn leads to enhanced sample-to-sample fluctuations and a larger variance. Meanwhile, the slight decrease of \mathbb{E}\left[E_{J}\right] is associated with finite-size effect. For a junction with a finite cross section, increasing \xi reduces the number of independent segments across the interface. Consequently, the probability of realizing rare, exceptionally thin barrier regions that would otherwise dominate the tunneling contribution is reduced, resulting in a modest suppression of \mathbb{E}\left[E_{J}\right] . When \xi approaches the junction dimension, the effective number of independent segments becomes O\left(1\right) , and both the mean and variance of E_{J} become insensitive to further increases in \xi , resulting in saturation behavior. Table 1: Josephson energy E_{J}/h (GHz) as a function of RMS roughness \sigma and transverse correlation length \xi . Values are given as mean \pm standard deviation. \xi (nm) \sigma=0.08 nm \sigma=0.09 nm \sigma=0.10 nm 10 17.92\pm 1.70 23.57\pm 3.20 32.32\pm 6.15 20 17.86\pm 3.05 23.39\pm 5.45 31.81\pm 9.82 30 17.72\pm 3.99 23.05\pm 6.87 31.06\pm 11.94 40 17.54\pm 4.70 22.67\pm 7.89 30.27\pm 13.35 50 17.37\pm 5.24 22.32\pm 8.64 29.56\pm 14.35 V Summary To summarize, we have investigated the variability of the Josephson energy E_{J} in Al/AlO {}_{\text{x}} /Al junctions arising from interface roughness. The roughness is modeled using a Gaussian random field characterized by the RMS amplitude \sigma and the transverse correlation length \xi . The values of these parameters are taken from the experimental literature and extracted from our atomistic MD simulations. The quantum transport model is based on the Ambegaokar–Baratoff relation combined with a local thickness approximation, both of which are justified and numerically verified for the Josephson junctions. Numerical simulations over an ensemble of 5,000 samples show that E_{J} follows a log-normal distribution. The mean value of E_{J} increases with \sigma and decreases slightly with \xi , while the variance of E_{J} increases with both \sigma and \xi . In this work, we have assumed that \sigma and \xi are independent structural parameters and are identical for the two interfaces in Al/AlO {}_{\text{x}} /Al junctions. While reasonable, these assumptions require careful experimental verification under different growth and process conditions. Connecting our model to the experimental characterization of the fabricated Josephson junctions and combining the superconductivity solvers employed here with computational microwave engineering software 11 will be important directions for future work, opening the door to the co-design of Josephson junctions and their neighboring superconducting-circuit elements. VI